Instrumental Variables Estimation without Outside Instruments

This paper considers an alternative estimation approach of regression models with endogenous regressors when external instruments are not available. An artificial neural network is used to model the correlation between error and regressors coupled with Bayesian exponentially tilted empirical likelihood to obtain a consistent estimation of the model's parameters. Monte Carlo simulations indicate that the new approach performs well in finite samples. An empirical application is presented to illustrate the usefulness of our proposed approach.